Add microlite_op_resolver to allow for using the all or mutable op resolver

[mocleiri]

I found out that there is a memory overhead for using the all op resolver. I'm not sure on the exact overhead but I head it could be 4k.

Lets create a new type to represent the op_resolver. It should support using the all resolver or the mutable op resolver where you can specify which one to load.

[mocleiri]

There are 74 ops loaded in the all op resolver but 82 ops listed in the mutable op resolver. I'm not sure why there is this disparity.

There are 146 ops listed at the tensorflow lite level so I created my own enumeration to represent the 82 possibilities.